When the pause disappears

September 2026

You get used to the little pause.

You ask an AI something. The words start appearing. Maybe you read them as they arrive. Maybe you check another tab. Maybe you are already thinking about the next question.

It is quick, but there is still a rhythm to it.

You ask.

It thinks.

It answers.

Then I tried ChatJimmy.

Taalas has built a chip specifically around Llama 3.1 8B. Its public demo runs at somewhere around 14,000–17,000 tokens a second. AMD has now agreed to acquire the company.

Fourteen thousand tokens a second is a slightly meaningless number until you actually experience it.

Then it makes sense.

The answer is just there.

Not gradually appearing. Not fast typing. Just there before you have really settled into waiting for it.

And I found that surprisingly disorientating.

Not because the AI was suddenly more intelligent.

Because something about the interaction had changed.

The pause had disappeared.

I hadn’t realised what the pause was doing

I have been thinking about this ever since.

At the speeds we are used to today, humans can still sort of keep up with AI.

Not literally, obviously. But conversationally.

You can watch what it is doing. Read something. Interrupt. Ask another question. Change direction. Approve the next step.

You are still, in some sense, pacing the conversation.

That becomes much more interesting when you stop thinking about one person chatting to one AI and start thinking about agents.

I already run a little studio of bots.

They pass work between themselves. One researches something. Another challenges it. Another writes. Another checks. Conversations happen that I am not directly part of.

At today’s speeds that can already feel a bit strange.

Now imagine those conversations happening at 14,000 tokens a second.

The human is not really in that conversation any more.

At least, not turn by turn.

And perhaps we shouldn’t be

My first reaction was: humans become the bottleneck.

Which I think is true.

But I am not sure that is quite the right way to think about it.

Because why would we try to keep up?

If two agents can resolve something between themselves in milliseconds, making them wait for me to read every exchange rather defeats the point.

The mistake might be trying to preserve the way we currently work.

AI does something.

Human checks it.

AI does the next thing.

Human checks it again.

As the machines get faster, that model starts to look increasingly odd.

But there is a problem.

The bit I don’t want to remove is human judgement.

An agent can research, draft, challenge another agent and pass work onwards.

It still cannot own the trust somebody has placed in me.

It cannot decide what I am prepared to stand behind.

And it cannot be accountable when the answer matters.

So perhaps speed does not remove the human from the system.

It changes where the human needs to sit.

Where does the judgement go?

That is the bit I am interested in now.

Maybe our job is increasingly not to participate in every conversation.

Maybe it is to decide what the agents are trying to achieve. What boundaries they operate within. What evidence matters. What requires escalation. And where a human absolutely has to make the call.

In other words, perhaps judgement moves from being continuously in the conversation to being deliberately placed at particular points around it.

I don’t have a neat framework for that yet.

But ChatJimmy made the question feel much less theoretical.

You can argue about how quickly AI will improve.

You can argue about agents.

You can argue about whether any of this will really change professional work.

Then you experience inference at that speed and something becomes quite tangible.

The machines are not going to need to wait for us.

So if human judgement is still the part we cannot outsource, the interesting question becomes:

Where exactly do we put it?

What this makes possible

I wrote more about where judgement still sits in Where is my place in a world of AI?. A practical overnight test of leaving work with agents is in Personal Assistants Are Starting to Just Work.